VAE-Based Interpretable Latent Variable Model for Process Monitoring.

Area of Science:

• Chemical Engineering
• Data Science
• Artificial Intelligence

Background:

• Traditional process monitoring (PM) relies on shallow learning, offering interpretability but limited performance.
• Deep learning (DL) excels in performance for PM but often lacks human-friendly interpretability.
• Designing interpretable DL-based latent variable models (LVMs) for PM remains a challenge.

Purpose of the Study:

• To develop a deep learning-based interpretable latent variable model for process monitoring.
• To enhance the interpretability of DL models in process monitoring applications.
• To improve fault detection accuracy and reduce data requirements in process monitoring.

Main Methods:

• A variational autoencoder-based interpretable LVM (VAE-ILVM) was developed.
• Taylor expansions guided the design of activation functions for VAE-ILVM.
• A de la Peña inequality was employed for threshold learning, treating exceedance counts as a martingale.

Main Results:

• The VAE-ILVM allows for non-disappearing fault impact terms in monitoring metrics.
• The proposed threshold learning method significantly reduces the minimum required sample size.
• Effectiveness was validated using two chemical process examples.

Conclusions:

• The VAE-ILVM offers a novel approach to interpretable deep learning for process monitoring.
• The integration of de la Peña inequality enhances model efficiency by reducing sample size requirements.
• This method provides a promising direction for developing more transparent and data-efficient PM systems.